The early detection of Brown Spot disease (Bipolaris oryzae) in tropical rice crops is hindered by the phenotypic ambiguity of early-stage lesions and the complexity of background noise in open-field canopies. Traditional monitoring methods often fail due to the spectral saturation of structural indices in dense vegetation. This study proposes a novel framework integrating proximal multispectral sensing (h = 40–50 cm) with a modernized deep learning architecture. Through a high-throughput audit of 17 multispectral indices, we engineered a high-dimensional Bio-Spectral Tensor fusing the Chlorophyll Vegetation Index (CVI), Modified Anthocyanin Content Index (MACI), and Green Normalized Difference Vegetation Index (GNDVI).Statistical analysis via the Fisher Discriminant Criterion revealed that CVI provides a 144-fold increase in class-discriminative power compared to the industry-standard NDVI, effectively decoupling biotic stress signals from environmental artifacts. These tensors were processed using the ConvNeXt architecture, a modernized convolutional model optimized for high-dimensional feature extraction. The system achieved an overall classification accuracy of 98.0% and a Recall of 1.00 for the diseased class, ensuring zero false negatives in phytosanitary monitoring. Furthermore, Explainable AI (Grad-CAM) analysis validated that the network's decision-making is driven by physiological lesion signatures—specifically necrotic centers and chlorotic halos—rather than background noise. These findings establish a scalable, physics-informed baseline for autonomous crop health monitoring in the Colombian tropics.